most projects focus on text line detection,this project use yolo2 to detect character which is easy for recognition

we use smooth L1 loss and cross entropy insted of L2 loss in darknet.
With smaller backbone,our speed could reach 300FPS

tensorflow 1.8(low version should also be able to run)
opencv

1.run train_from_Synthetic.py
2.run test.py

1.put your image in data/image
2.put your label in data/label
3.run train.py